Cross-Supervision Similarity Network for Medical Image Classification on Imbalanced Small Datasets

Area of Science:

• Machine Learning
• Medical Imaging Analysis
• Computer Vision

Background:

• Imbalanced small datasets are prevalent in medical imaging machine learning, posing challenges for reliable classification.
• Existing generative methods struggle with subtle distinctions in medical images, while few-shot learning requires balanced data.
• Accurate classification is crucial for medical applications like cancer prognosis using MRIs.

Purpose of the Study:

• To propose a novel network, Cross-Supervision Similarity Network (CSSN), for robust classification on imbalanced small medical imaging datasets.
• To address limitations of generative and few-shot learning methods in medical image analysis.
• To enhance the reliability of machine learning models in clinical applications.

Main Methods:

• Developed a Cross-Supervision Similarity Network (CSSN) utilizing cross-supervision between class and patch features.
• Transformed classification into a comparison task by calculating similarity scores at patch and class scales.
• Employed soft logarithmic supervision to balance training difficulty between similarity branches.

Main Results:

• CSSN achieved significant performance improvements: 15% F1 score, 6% accuracy, and 9% balanced accuracy on the PCR-ISD dataset.
• Demonstrated superiority in identifying minority classes and enhancing classification capabilities.
• Validated effectiveness and generalization ability through extensive experiments on three datasets.

Conclusions:

• CSSN offers a superior solution for classification tasks involving imbalanced small datasets in medical imaging.
• The method effectively handles challenging medical image classification scenarios, such as pCR evaluation.
• The proposed approach shows strong potential for improving machine learning reliability in clinical practice.